Methodology

Sampling comes first

Every ingest returns count, median interval, interval MAD, coverage, duplicates, and explicit gaps. The primary periodogram never interpolates.

Generalized Lomb-Scargle

At each angular frequency, telespec solves the weighted least-squares model y(t) = a cos(wt) + b sin(wt) + c. Power is the fractional chi-square reduction from the floating-mean null. This is the generalized form in Zechmeister and Kuerster (2009), not the zero-mean classical periodogram.

The global false-alarm threshold is the selected quantile of maximum periodogram power after repeatedly permuting values over fixed timestamps. This controls the multiple-frequency search empirically.

Localization and detection

STFT and Morlet CWT require a regular grid, so that resampling is explicit and rejected across large gaps. Sliding-window log octave-band powers and remainder variance feed a robust Minimum Covariance Determinant baseline. Two-sided CUSUM detects persistent level changes. Findings retain interval, magnitude, confidence, and evidence.

References: Zechmeister & Kuerster, A&A 496 (2009); Baluev, MNRAS 385 (2008); Page, Biometrika 41 (1954); Rousseeuw & Van Driessen, Technometrics 41 (1999).